The machine drives the CLI programmatically rather than through a terminal, so there is no ANSI to scrape and no screen to parse. When the agent needs a human, the request arrives as structured data — and the console can render the thing itself.
- A file edit arrives as a real inline diff.
- A shell command arrives as the command.
- A plan review arrives as rendered Markdown.
- A question arrives as option cards, with radio buttons or checkboxes.
Four ways to answer
Allow once. Always allow — which surfaces the CLI’s own standing-grant offer, so you are not re-approving the same safe thing forever. Deny. Or deny with a reason, which the model reads and works around rather than simply stopping.
There is also Cancel, which stops the turn outright: the Escape key that a web console otherwise does not have.
It stays in the record
Answers remain visible in the transcript afterwards. What was asked, what was decided and who decided it are part of the conversation rather than a modal that vanished.